Google bets $17 billion on Wyoming as datacentres fight for power and cooling

Wyoming, at $17 billion

Google is building its first data centre in Wyoming, backed by a $17 billion investment. The facility will feature a 702 MW fuel cell development integrated with carbon capture and storage, developed in partnership with American Electric Power and Related Digital among others. The project is expected to generate hundreds of millions of dollars in local property taxes for Laramie County and School District 1, and Google and its partners are establishing a Community Council to support Cheyenne's growth. m:…2698

The engineering choices are the story. The Wyoming data centre will use air cooling to minimise strain on local water resources — a direct answer to the water politics that now follow every large data centre — and carbon emissions will be stored at Tallgrass's Wyoming Sequestration Hub, with permanent geologic carbon storage. Google's total contracted carbon-capture capacity is nearly 1 gigawatt. A data centre that generates its own power from fuel cells, captures the carbon, and avoids the local water supply is a different kind of facility from the ones that defined the last decade, and it is also a much more expensive one to build. m:…2698

Nautilus Data Technologies, presenting at DCD Cooling Week 2026 on November 11, shows how much of the bill is decided by small numbers. Cooling distribution unit approach temperature determines how warm facility water can run while meeting chip cooling requirements, and removing one degree of approach temperature allows plants to run on free cooling at warmer outdoor temperatures. Reducing approach temperature to 2°C adds free-cooling hours throughout the year, and the company's EcoCore FCD unit is rated down to a 2°C approach. m:…6699

Two degrees is worth 480 hours

Nautilus modelling for Ashburn, Virginia puts a number on it: 480 extra free-cooling hours per year for a 2°C cooling unit compared with a 4°C unit. Fewer hours of mechanical chilling reduce power use and operating costs, and raising supply temperatures allows chip requirements to be met without overloading existing chillers. Different water sources could support year-round free cooling, with mechanical systems kept in reserve. The point is that cooling is not a fixed cost of running AI hardware; it is a design variable, and two degrees of approach temperature is worth roughly three weeks of free cooling a year in Virginia. m:…6699

Aria Networks makes the parallel argument for the network. Meridian Cloud has selected Aria as its networking partner to build standardised, inference-ready AI clusters for its dedicated inference cloud, scheduled to launch in Q2 2027. The network accounts for 10 to 15 percent of total AI cluster capital spending — a share large enough to decide whether a cluster makes money — and grey failures, such as microsecond congestion, can cause significant inference performance degradation. The two companies will develop standardised, validated network designs, with the first joint reference designs available to customers in Q2 2027. m:…2288

Aria's pitch is visibility. Its platform maximises token efficiency by detecting and resolving network-layer issues that traditional monitoring tools miss, its 800G switch platforms run AriaOS and capture telemetry at 100 to 10,000 times the resolution of traditional tools, and its forward-deployed engineers are embedded with the Meridian team to assist with network design and deployment planning. A network that cannot see microsecond congestion cannot fix it, and at inference scale, congestion is lost tokens. m:…2288

What it adds up to

Put the three stories together and the AI buildout looks less like a chip race and more like a civil engineering project. Seventeen billion dollars for one Wyoming site, 480 hours of free cooling from two degrees of approach temperature, and 10 to 15 percent of cluster capital spending on the network — these are the numbers that now decide whether AI capacity is profitable. The chip announcements will keep coming, but the binding constraints are power contracts, water, cooling design, and network telemetry. m:…2698 m:…2288

There is also a political lesson in the Wyoming announcement. A $17 billion project that pays hundreds of millions in local property taxes and sets up a Community Council is buying a social licence as well as a site. Data centres that arrive without one are finding their permits contested and their water use questioned. The cheapest megawatt is the one the neighbours do not object to, and Google's Wyoming design — air cooling, captured carbon, local taxes — is priced accordingly. m:…2698